Some perfomance thoughts required

[Alexander Adam]

Hi!

I've got a pretty complex data structure that keeps instances of
structs allocated with the new operator.
Now my issue is that those structures are required to be allocated /
deallocated pretty often resulting
in a huge perfomance issue. Usually what I am doing is to create
arrays and initialize their sizes a few
times bigger than originally required to avoid another allocation on
the next run. Then I do simply reset them
and do a memcpy at the existing location. This works great with POD
structus but my structs are complex
in the way that they not only use inheritance but they do also use
internal objects that do require an
constructor / destructor. So what is the general way to handle such
things? Am I missing something?

thanks!
Alex

 

[Erik Wikström]

Hi!

I've got a pretty complex data structure that keeps instances of
structs allocated with the new operator.
Now my issue is that those structures are required to be allocated /
deallocated pretty often resulting
in a huge perfomance issue. Usually what I am doing is to create
arrays and initialize their sizes a few
times bigger than originally required to avoid another allocation on
the next run. Then I do simply reset them
and do a memcpy at the existing location. This works great with POD
structus but my structs are complex
in the way that they not only use inheritance but they do also use
internal objects that do require an
constructor / destructor. So what is the general way to handle such
things? Am I missing something?

A pool allocator might help.

 

[tommy.hinks]

A pool allocator might help.

Check out boost:bject_pool<T>.

T

 

[Jerry Coffin]

Hi!

I've got a pretty complex data structure that keeps instances of
structs allocated with the new operator.
Now my issue is that those structures are required to be allocated /
deallocated pretty often resulting
in a huge perfomance issue. Usually what I am doing is to create
arrays and initialize their sizes a few
times bigger than originally required to avoid another allocation on
the next run. Then I do simply reset them
and do a memcpy at the existing location. This works great with POD
structus but my structs are complex
in the way that they not only use inheritance but they do also use
internal objects that do require an
constructor / destructor. So what is the general way to handle such
things? Am I missing something?

You probably want to overload operator new for the class(es) of the
objects you're allocating/freeing so frequently. At startup, your
operator new will grab a big chunk of memory for the data, and break it
up into chunks, each the size of a single object. As objects are
allocated, it'll give out the addresses of object-sized chunks, and mark
each chunk as being in use when it does so. When an object is freed,
it'll mark each chunk as being available again.

You can probably find existing source code for a few different versions
of this -- Andrei Alexandrescu's Loki library has one, and Boost has
another, and so on.

 

[Fei Liu]

Hi!

I've got a pretty complex data structure that keeps instances of
structs allocated with the new operator.
Now my issue is that those structures are required to be allocated /
deallocated pretty often resulting
in a huge perfomance issue. Usually what I am doing is to create
arrays and initialize their sizes a few
times bigger than originally required to avoid another allocation on
the next run. Then I do simply reset them
and do a memcpy at the existing location. This works great with POD
structus but my structs are complex
in the way that they not only use inheritance but they do also use
internal objects that do require an
constructor / destructor. So what is the general way to handle such
things? Am I missing something?

thanks!
Alex

Apart from the other suggestions, you should think about the use of
array as your container data structure, perhaps a node based container
will provide better overall performance, especially when it's memory
intensive.

Fei